Xinyu Zhang

Sifan Zhou, Zhi Tian, Xiangxiang Chu et al.

The deployment of 3D detectors strikes one of the major challenges in real-world self-driving scenarios. Existing BEV-based (i.e., Bird Eye View) detectors favor sparse convolutions (known as SPConv) to speed up training and inference, which puts a hard barrier for deployment, especially for on-device applications. In this paper, to tackle the challenge of efficient 3D object detection from an industry perspective, we devise a deployment-friendly pillar-based 3D detector, termed FastPillars. First, we introduce a novel lightweight Max-and-Attention Pillar Encoding (MAPE) module specially for enhancing small 3D objects. Second, we propose a simple yet effective principle for designing a backbone in pillar-based 3D detection. We construct FastPillars based on these designs, achieving high performance and low latency without SPConv. Extensive experiments on two large-scale datasets demonstrate the effectiveness and efficiency of FastPillars for on-device 3D detection regarding both performance and speed. Specifically, FastPillars delivers state-of-the-art accuracy on Waymo Open Dataset with 1.8X speed up and 3.8 mAPH/L2 improvement over CenterPoint (SPConv-based). Our code is publicly available at: https://github.com/StiphyJay/FastPillars.

Lei Yang, Xinyu Zhang, Li Wang et al.

Monocular 3D object detection is an essential perception task for autonomous driving. However, the high reliance on large-scale labeled data make it costly and time-consuming during model optimization. To reduce such over-reliance on human annotations, we propose Mix-Teaching, an effective semi-supervised learning framework applicable to employ both labeled and unlabeled images in training stage. Mix-Teaching first generates pseudo-labels for unlabeled images by self-training. The student model is then trained on the mixed images possessing much more intensive and precise labeling by merging instance-level image patches into empty backgrounds or labeled images. This is the first to break the image-level limitation and put high-quality pseudo labels from multi frames into one image for semi-supervised training. Besides, as a result of the misalignment between confidence score and localization quality, it's hard to discriminate high-quality pseudo-labels from noisy predictions using only confidence-based criterion. To that end, we further introduce an uncertainty-based filter to help select reliable pseudo boxes for the above mixing operation. To the best of our knowledge, this is the first unified SSL framework for monocular 3D object detection. Mix-Teaching consistently improves MonoFlex and GUPNet by significant margins under various labeling ratios on KITTI dataset. For example, our method achieves around +6.34% AP@0.7 improvement against the GUPNet baseline on validation set when using only 10% labeled data. Besides, by leveraging full training set and the additional 48K raw images of KITTI, it can further improve the MonoFlex by +4.65% improvement on AP@0.7 for car detection, reaching 18.54% AP@0.7, which ranks the 1st place among all monocular based methods on KITTI test leaderboard. The code and pretrained models will be released at https://github.com/yanglei18/Mix-Teaching.

Xinyu Zhang, Li Wang, Jian Chen et al.

Radar has stronger adaptability in adverse scenarios for autonomous driving environmental perception compared to widely adopted cameras and LiDARs. Compared with commonly used 3D radars, the latest 4D radars have precise vertical resolution and higher point cloud density, making it a highly promising sensor for autonomous driving in complex environmental perception. However, due to the much higher noise than LiDAR, manufacturers choose different filtering strategies, resulting in an inverse ratio between noise level and point cloud density. There is still a lack of comparative analysis on which method is beneficial for deep learning-based perception algorithms in autonomous driving. One of the main reasons is that current datasets only adopt one type of 4D radar, making it difficult to compare different 4D radars in the same scene. Therefore, in this paper, we introduce a novel large-scale multi-modal dataset featuring, for the first time, two types of 4D radars captured simultaneously. This dataset enables further research into effective 4D radar perception algorithms.Our dataset consists of 151 consecutive series, most of which last 20 seconds and contain 10,007 meticulously synchronized and annotated frames. Moreover, our dataset captures a variety of challenging driving scenarios, including many road conditions, weather conditions, nighttime and daytime with different lighting intensities and periods. Our dataset annotates consecutive frames, which can be applied to 3D object detection and tracking, and also supports the study of multi-modal tasks. We experimentally validate our dataset, providing valuable results for studying different types of 4D radars. This dataset is released on https://github.com/adept-thu/Dual-Radar.

Xinyu Zhang, Jiahui Chen, Junkun Yuan et al. · tencent-ai

Masked image modeling (MIM) learns visual representation by masking and reconstructing image patches. Applying the reconstruction supervision on the CLIP representation has been proven effective for MIM. However, it is still under-explored how CLIP supervision in MIM influences performance. To investigate strategies for refining the CLIP-targeted MIM, we study two critical elements in MIM, i.e., the supervision position and the mask ratio, and reveal two interesting perspectives, relying on our developed simple pipeline, context autodecoder with CLIP target (CAE v2). Firstly, we observe that the supervision on visible patches achieves remarkable performance, even better than that on masked patches, where the latter is the standard format in the existing MIM methods. Secondly, the optimal mask ratio positively correlates to the model size. That is to say, the smaller the model, the lower the mask ratio needs to be. Driven by these two discoveries, our simple and concise approach CAE v2 achieves superior performance on a series of downstream tasks. For example, a vanilla ViT-Large model achieves 81.7% and 86.7% top-1 accuracy on linear probing and fine-tuning on ImageNet-1K, and 55.9% mIoU on semantic segmentation on ADE20K with the pre-training for 300 epochs. We hope our findings can be helpful guidelines for the pre-training in the MIM area, especially for the small-scale models.

Teng Xi, Yifan Sun, Deli Yu et al.

This paper proposes a novel Unified Feature Optimization (UFO) paradigm for training and deploying deep models under real-world and large-scale scenarios, which requires a collection of multiple AI functions. UFO aims to benefit each single task with a large-scale pretraining on all tasks. Compared with the well known foundation model, UFO has two different points of emphasis, i.e., relatively smaller model size and NO adaptation cost: 1) UFO squeezes a wide range of tasks into a moderate-sized unified model in a multi-task learning manner and further trims the model size when transferred to down-stream tasks. 2) UFO does not emphasize transfer to novel tasks. Instead, it aims to make the trimmed model dedicated for one or more already-seen task. With these two characteristics, UFO provides great convenience for flexible deployment, while maintaining the benefits of large-scale pretraining. A key merit of UFO is that the trimming process not only reduces the model size and inference consumption, but also even improves the accuracy on certain tasks. Specifically, UFO considers the multi-task training and brings two-fold impact on the unified model: some closely related tasks have mutual benefits, while some tasks have conflicts against each other. UFO manages to reduce the conflicts and to preserve the mutual benefits through a novel Network Architecture Search (NAS) method. Experiments on a wide range of deep representation learning tasks (i.e., face recognition, person re-identification, vehicle re-identification and product retrieval) show that the model trimmed from UFO achieves higher accuracy than its single-task-trained counterpart and yet has smaller model size, validating the concept of UFO. Besides, UFO also supported the release of 17 billion parameters computer vision (CV) foundation model which is the largest CV model in the industry.

Yu Wang, Xiangbo Su, Qiang Chen et al.

Open-vocabulary object detection focusing on detecting novel categories guided by natural language. In this report, we propose Open-Vocabulary Light-Weighted Detection Transformer (OVLW-DETR), a deployment friendly open-vocabulary detector with strong performance and low latency. Building upon OVLW-DETR, we provide an end-to-end training recipe that transferring knowledge from vision-language model (VLM) to object detector with simple alignment. We align detector with the text encoder from VLM by replacing the fixed classification layer weights in detector with the class-name embeddings extracted from the text encoder. Without additional fusing module, OVLW-DETR is flexible and deployment friendly, making it easier to implement and modulate. improving the efficiency of interleaved attention computation. Experimental results demonstrate that the proposed approach is superior over existing real-time open-vocabulary detectors on standard Zero-Shot LVIS benchmark. Source code and pre-trained models are available at [https://github.com/Atten4Vis/LW-DETR].

Xinyu Zhang, Zhiwei Li, Zhenhong Zou et al. · tsinghua

Noise has always been nonnegligible trouble in object detection by creating confusion in model reasoning, thereby reducing the informativeness of the data. It can lead to inaccurate recognition due to the shift in the observed pattern, that requires a robust generalization of the models. To implement a general vision model, we need to develop deep learning models that can adaptively select valid information from multi-modal data. This is mainly based on two reasons. Multi-modal learning can break through the inherent defects of single-modal data, and adaptive information selection can reduce chaos in multi-modal data. To tackle this problem, we propose a universal uncertainty-aware multi-modal fusion model. It adopts a multi-pipeline loosely coupled architecture to combine the features and results from point clouds and images. To quantify the correlation in multi-modal information, we model the uncertainty, as the inverse of data information, in different modalities and embed it in the bounding box generation. In this way, our model reduces the randomness in fusion and generates reliable output. Moreover, we conducted a completed investigation on the KITTI 2D object detection dataset and its derived dirty data. Our fusion model is proven to resist severe noise interference like Gaussian, motion blur, and frost, with only slight degradation. The experiment results demonstrate the benefits of our adaptive fusion. Our analysis on the robustness of multi-modal fusion will provide further insights for future research.

Lei Yang, Jiaxin Yu, Xinyu Zhang et al.

Although the majority of recent autonomous driving systems concentrate on developing perception methods based on ego-vehicle sensors, there is an overlooked alternative approach that involves leveraging intelligent roadside cameras to help extend the ego-vehicle perception ability beyond the visual range. We discover that most existing monocular 3D object detectors rely on the ego-vehicle prior assumption that the optical axis of the camera is parallel to the ground. However, the roadside camera is installed on a pole with a pitched angle, which makes the existing methods not optimal for roadside scenes. In this paper, we introduce a novel framework for Roadside Monocular 3D object detection with ground-aware embeddings, named MonoGAE. Specifically, the ground plane is a stable and strong prior knowledge due to the fixed installation of cameras in roadside scenarios. In order to reduce the domain gap between the ground geometry information and high-dimensional image features, we employ a supervised training paradigm with a ground plane to predict high-dimensional ground-aware embeddings. These embeddings are subsequently integrated with image features through cross-attention mechanisms. Furthermore, to improve the detector's robustness to the divergences in cameras' installation poses, we replace the ground plane depth map with a novel pixel-level refined ground plane equation map. Our approach demonstrates a substantial performance advantage over all previous monocular 3D object detectors on widely recognized 3D detection benchmarks for roadside cameras. The code and pre-trained models will be released soon.

Li Wang, Xinyu Zhang, Fachuan Zhao et al.

Non-maximum suppression (NMS) is an essential post-processing module used in many 3D object detection frameworks to remove overlapping candidate bounding boxes. However, an overreliance on classification scores and difficulties in determining appropriate thresholds can affect the resulting accuracy directly. To address these issues, we introduce fuzzy learning into NMS and propose a novel generalized Fuzzy-NMS module to achieve finer candidate bounding box filtering. The proposed Fuzzy-NMS module combines the volume and clustering density of candidate bounding boxes, refining them with a fuzzy classification method and optimizing the appropriate suppression thresholds to reduce uncertainty in the NMS process. Adequate validation experiments are conducted using the mainstream KITTI and large-scale Waymo 3D object detection benchmarks. The results of these tests demonstrate the proposed Fuzzy-NMS module can improve the accuracy of numerous recently NMS-based detectors significantly, including PointPillars, PV-RCNN, and IA-SSD, etc. This effect is particularly evident for small objects such as pedestrians and bicycles. As a plug-and-play module, Fuzzy-NMS does not need to be retrained and produces no obvious increases in inference time.

Xinyu Zhang, Li Wang, Zhiqiang Jiang et al.

Local Feature Matching, an essential component of several computer vision tasks (e.g., structure from motion and visual localization), has been effectively settled by Transformer-based methods. However, these methods only integrate long-range context information among keypoints with a fixed receptive field, which constrains the network from reconciling the importance of features with different receptive fields to realize complete image perception, hence limiting the matching accuracy. In addition, these methods utilize a conventional handcrafted encoding approach to integrate the positional information of keypoints into the visual descriptors, which limits the capability of the network to extract reliable positional encoding message. In this study, we propose Feature Matching with Reconciliatory Transformer (FMRT), a novel Transformer-based detector-free method that reconciles different features with multiple receptive fields adaptively and utilizes parallel networks to realize reliable positional encoding. Specifically, FMRT proposes a dedicated Reconciliatory Transformer (RecFormer) that consists of a Global Perception Attention Layer (GPAL) to extract visual descriptors with different receptive fields and integrate global context information under various scales, Perception Weight Layer (PWL) to measure the importance of various receptive fields adaptively, and Local Perception Feed-forward Network (LPFFN) to extract deep aggregated multi-scale local feature representation. Extensive experiments demonstrate that FMRT yields extraordinary performance on multiple benchmarks, including pose estimation, visual localization, homography estimation, and image matching.

Lei Yang, Xinyu Zhang, Jun Li et al.

Roadside perception can greatly increase the safety of autonomous vehicles by extending their perception ability beyond the visual range and addressing blind spots. However, current state-of-the-art vision-based roadside detection methods possess high accuracy on labeled scenes but have inferior performance on new scenes. This is because roadside cameras remain stationary after installation and can only collect data from a single scene, resulting in the algorithm overfitting these roadside backgrounds and camera poses. To address this issue, in this paper, we propose an innovative Scenario Generalization Framework for Vision-based Roadside 3D Object Detection, dubbed SGV3D. Specifically, we employ a Background-suppressed Module (BSM) to mitigate background overfitting in vision-centric pipelines by attenuating background features during the 2D to bird's-eye-view projection. Furthermore, by introducing the Semi-supervised Data Generation Pipeline (SSDG) using unlabeled images from new scenes, diverse instance foregrounds with varying camera poses are generated, addressing the risk of overfitting specific camera poses. We evaluate our method on two large-scale roadside benchmarks. Our method surpasses all previous methods by a significant margin in new scenes, including +42.57% for vehicle, +5.87% for pedestrian, and +14.89% for cyclist compared to BEVHeight on the DAIR-V2X-I heterologous benchmark. On the larger-scale Rope3D heterologous benchmark, we achieve notable gains of 14.48% for car and 12.41% for large vehicle. We aspire to contribute insights on the exploration of roadside perception techniques, emphasizing their capability for scenario generalization. The code will be available at https://github.com/yanglei18/SGV3D

Yingshan Liang, Keyu Fan, Zhicheng Du et al.

Video-to-audio (V2A) generation shows great potential in fields such as film production. Despite significant advances, current V2A methods relying on global video information struggle with complex scenes and generating audio tailored to specific objects. To address these limitations, we introduce Hear-Your-Click, an interactive V2A framework enabling users to generate sounds for specific objects by clicking on the frame. To achieve this, we propose Object-aware Contrastive Audio-Visual Fine-tuning (OCAV) with a Mask-guided Visual Encoder (MVE) to obtain object-level visual features aligned with audio. Furthermore, we tailor two data augmentation strategies, Random Video Stitching (RVS) and Mask-guided Loudness Modulation (MLM), to enhance the model's sensitivity to segmented objects. To measure audio-visual correspondence, we designed a new evaluation metric, the CAV score. Extensive experiments demonstrate that our framework offers more precise control and improves generation performance across various metrics. Project Page: https://github.com/SynapGrid/Hear-Your-Click

Qiang Chen, Xiangbo Su, Xinyu Zhang et al.

In this paper, we present a light-weight detection transformer, LW-DETR, which outperforms YOLOs for real-time object detection. The architecture is a simple stack of a ViT encoder, a projector, and a shallow DETR decoder. Our approach leverages recent advanced techniques, such as training-effective techniques, e.g., improved loss and pretraining, and interleaved window and global attentions for reducing the ViT encoder complexity. We improve the ViT encoder by aggregating multi-level feature maps, and the intermediate and final feature maps in the ViT encoder, forming richer feature maps, and introduce window-major feature map organization for improving the efficiency of interleaved attention computation. Experimental results demonstrate that the proposed approach is superior over existing real-time detectors, e.g., YOLO and its variants, on COCO and other benchmark datasets. Code and models are available at (https://github.com/Atten4Vis/LW-DETR).

Huanan Wang, Xinyu Zhang, Jun Li et al.

Due to the high complexity and occlusion, insufficient perception in the crowded urban intersection can be a serious safety risk for both human drivers and autonomous algorithms, whereas CVIS (Cooperative Vehicle Infrastructure System) is a proposed solution for full-participants perception in this scenario. However, the research on roadside multimodal perception is still in its infancy, and there is no open-source dataset for such scenario. Accordingly, this paper fills the gap. Through an IPS (Intersection Perception System) installed at the diagonal of the intersection, this paper proposes a high-quality multimodal dataset for the intersection perception task. The center of the experimental intersection covers an area of 3000m2, and the extended distance reaches 300m, which is typical for CVIS. The first batch of open-source data includes 14198 frames, and each frame has an average of 319.84 labels, which is 9.6 times larger than the most crowded dataset (H3D dataset in 2019) by now. In order to facilitate further study, this dataset tries to keep the label documents consistent with the KITTI dataset, and a standardized benchmark is created for algorithm evaluation. Our dataset is available at: http://www.openmpd.com/column/other_datasets.

Lei Yang, Xinyu Zhang, Li Wang et al.

3D object detection with a single image is an essential and challenging task for autonomous driving. Recently, keypoint-based monocular 3D object detection has made tremendous progress and achieved great speed-accuracy trade-off. However, there still exists a huge gap with LIDAR-based methods in terms of accuracy. To improve their performance without sacrificing efficiency, we propose a sort of lightweight feature pyramid network called Lite-FPN to achieve multi-scale feature fusion in an effective and efficient way, which can boost the multi-scale detection capability of keypoint-based detectors. Besides, the misalignment between classification score and localization precision is further relieved by introducing a novel regression loss named attention loss. With the proposed loss, predictions with high confidence but poor localization are treated with more attention during the training phase. Comparative experiments based on several state-of-the-art keypoint-based detectors on the KITTI dataset show that our proposed methods manage to achieve significant improvements in both accuracy and frame rate. The code and pretrained models will be released at \url{https://github.com/yanglei18/Lite-FPN}.

Yanpeng Sun, Huaxin Zhang, Qiang Chen et al.

We focus on improving the visual understanding capability for boosting the vision-language models. We propose \textbf{Arcana}, a multiModal language model, which introduces two crucial techniques. First, we present Multimodal LoRA (MM-LoRA), a module designed to enhance the decoder. Unlike traditional language-driven decoders, MM-LoRA consists of two parallel LoRAs -- one for vision and one for language -- each with its own parameters. This disentangled parameters design allows for more specialized learning in each modality and better integration of multimodal information. Second, we introduce the Query Ladder adapter (QLadder) to improve the visual encoder. QLadder employs a learnable ``\textit{ladder}'' structure to deeply aggregates the intermediate representations from the frozen pretrained visual encoder (e.g., CLIP image encoder). This enables the model to learn new and informative visual features, as well as remaining the powerful capabilities of the pretrained visual encoder. These techniques collectively enhance Arcana's visual perception power, enabling it to leverage improved visual information for more accurate and contextually relevant outputs across various multimodal scenarios. Extensive experiments and ablation studies demonstrate the effectiveness and generalization capability of our Arcana. The code and re-annotated data are available at \url{https://arcana-project-page.github.io}.

Chenyu Zheng, Xinyu Zhang, Rongzhen Wang et al.

Diffusion Transformers have emerged as the foundation for vision generative models, but their scalability is limited by the high cost of hyperparameter (HP) tuning at large scales. Recently, Maximal Update Parametrization ($μ$P) was proposed for vanilla Transformers, which enables stable HP transfer from small to large language models, and dramatically reduces tuning costs. However, it remains unclear whether $μ$P of vanilla Transformers extends to diffusion Transformers, which differ architecturally and objectively. In this work, we generalize standard $μ$P to diffusion Transformers and validate its effectiveness through large-scale experiments. First, we rigorously prove that $μ$P of mainstream diffusion Transformers, including U-ViT, DiT, PixArt-$α$, and MMDiT, aligns with that of the vanilla Transformer, enabling the direct application of existing $μ$P methodologies. Leveraging this result, we systematically demonstrate that DiT-$μ$P enjoys robust HP transferability. Notably, DiT-XL-2-$μ$P with transferred learning rate achieves 2.9 times faster convergence than the original DiT-XL-2. Finally, we validate the effectiveness of $μ$P on text-to-image generation by scaling PixArt-$α$ from 0.04B to 0.61B and MMDiT from 0.18B to 18B. In both cases, models under $μ$P outperform their respective baselines while requiring small tuning cost, only 5.5% of one training run for PixArt-$α$ and 3% of consumption by human experts for MMDiT-18B. These results establish $μ$P as a principled and efficient framework for scaling diffusion Transformers.

Zhenyu Xu, Xiaoqi Shen, Haotian Nan et al.

Instruction-based image editing enables intuitive manipulation through natural language commands. However, text instructions alone often lack the precision required for fine-grained control over edit intensity. We introduce NumeriKontrol, a framework that allows users to precisely adjust image attributes using continuous scalar values with common units. NumeriKontrol encodes numeric editing scales via an effective Numeric Adapter and injects them into diffusion models in a plug-and-play manner. Thanks to a task-separated design, our approach supports zero-shot multi-condition editing, allowing users to specify multiple instructions in any order. To provide high-quality supervision, we synthesize precise training data from reliable sources, including high-fidelity rendering engines and DSLR cameras. Our Common Attribute Transform (CAT) dataset covers diverse attribute manipulations with accurate ground-truth scales, enabling NumeriKontrol to function as a simple yet powerful interactive editing studio. Extensive experiments show that NumeriKontrol delivers accurate, continuous, and stable scale control across a wide range of attribute editing scenarios. These contributions advance instruction-based image editing by enabling precise, scalable, and user-controllable image manipulation.

Li Wang, Chenfei Wang, Xinyu Zhang et al.

3D object detection plays a crucial role in environmental perception for autonomous vehicles, which is the prerequisite of decision and control. This paper analyses partition-based methods' inherent drawbacks. In the partition operation, a single instance such as a pedestrian is sliced into several pieces, which we call it the partition effect. We propose the Spatial-Attention Graph Convolution (S-AT GCN), forming the Feature Enhancement (FE) layers to overcome this drawback. The S-AT GCN utilizes the graph convolution and the spatial attention mechanism to extract local geometrical structure features. This allows the network to have more meaningful features for the foreground. Our experiments on the KITTI 3D object and bird's eye view detection show that S-AT Conv and FE layers are effective, especially for small objects. FE layers boost the pedestrian class performance by 3.62\% and cyclist class by 4.21\% 3D mAP. The time cost of these extra FE layers are limited. PointPillars with FE layers can achieve 48 PFS, satisfying the real-time requirement.